On 9/2/26 10:06, Gerd Hoffmann wrote:
>   Hi,
> 
>> Commit 7735ed4f8eb3afaf1a97d27a8862f382c387f0bc
>> ("OvmfPkg/PlatformInitLib: Set dynamic MMIO window size to 1/4")
>> breaks boot for both SEV-ES and SEV-SNP guests. The guest never reaches
>> the OS; firmware stops after binding the virtio-blk device.
> 
>> Found Mass Storage device: PciRoot(0x0)/Pci(0x2,0x0)/Pci(0x0,0x0)
>> InstallProtocolInterface: FA920010-6785-4941-B6EC-498C579F160A 7DD42020
> 
> Reproduces with any kind of virtio device.
> 
>> qemu-system-x86_64: Convert non guest_memfd backed memory region
>> (0x8040000000 ,+ 0x1000) to private
> 
>> 2026-08-27 03:16:02: MMIO using encrypted memory: 8040000014
> 
> Hmm.  So the patch moves the mmio window around and makes it larger:
> 
> -PlatformDynamicMmioWindow:   MMIO Space 0x2000000000 (128 GB)
> +PlatformDynamicMmioWindow:   MMIO Space 0x4000000000 (256 GB)
> 
> -PlatformDynamicMmioWindow:   Pci64 Base 0xC000000000
> -PlatformDynamicMmioWindow:   Pci64 Size 0x2000000000
> +PlatformDynamicMmioWindow:   Pci64 Base 0x8000000000
> +PlatformDynamicMmioWindow:   Pci64 Size 0x4000000000
> 
> In both cases the region is covered by an
> MemEncryptSevClearMmioPageEncMask call, so the encryption bit
> should not be set.  The error messages suggest it is set though.
> There is absolutely no difference in the logs beside the changed
> address and size.
> 
> Test branch with some debug logging is here:
> https://github.com/kraxel/edk2/commits/devel/sev-regression/
> 
> The branch also has patches explicitly tagging the 64-bit pci
> mmio window as mmio region.  Noticed this is not the case while
> looking into this.  That change doesn't make any difference in
> behavior though.
> 
> Tom?  Any clue what this might be?  Maybe something in the page
> table code?  IIRC you have a branch switching the home-grown
> sev page table code to use CpuPageTableLib instead, maybe this
> fixes the issue?

I rebased my CpuPageTableLib patch series on the main branch and it did
work.

I did some instrumenting and the range from 180000000 to 0xFB80000000 is
setup as a series of 1G pages all with the encryption bit cleared. When
I dump the pagetable entry at the moment before the crash I see that
particular entry has been split into 2M pages with the encryption bit
set - which leads to the issue.

It looks like there's a bug in SetMemoryEncDec() and how it tries to
mark page tables as read-only. Because the MMIO range now exceeds 512G,
the call to EnablePageTableProtection() isn't pointing at the base of
the page table and SetPageTablePoolReadOnly() is using an index value as
if it is.

I'll post a fix shortly. FYI, the reason why my CpuPageTableLib patch
series works is because I changed how the pagetable read-only protection
was done.

Thanks,
Tom

> 
> take care,
>   Gerd
> 



-=-=-=-=-=-=-=-=-=-=-=-
Groups.io Links: You receive all messages sent to this group.
View/Reply Online (#122160): https://edk2.groups.io/g/devel/message/122160
Mute This Topic: https://groups.io/mt/120969895/21656
Group Owner: [email protected]
Unsubscribe: https://edk2.groups.io/g/devel/unsub [[email protected]]
-=-=-=-=-=-=-=-=-=-=-=-


Reply via email to